ldiskfs/kernel_patches/patches/ubuntu2004/ext4-simple-blockalloc.patch

   1 diff -ur a/fs/ext4/ext4.h b/fs/ext4/ext4.h
   2 --- a/fs/ext4/ext4.h    2021-06-28 08:57:13.741381853 -0600
   3 +++ b/fs/ext4/ext4.h    2021-06-28 08:58:52.432392498 -0600
   4 @@ -1519,6 +1519,9 @@
   5         unsigned int s_mb_min_to_scan;
   6         unsigned int s_mb_stats;
   7         unsigned int s_mb_order2_reqs;
   8 +       ext4_fsblk_t s_mb_c1_blocks;
   9 +       ext4_fsblk_t s_mb_c2_blocks;
  10 +       ext4_fsblk_t s_mb_c3_blocks;
  11         unsigned long *s_mb_prealloc_table;
  12         unsigned int s_mb_group_prealloc;
  13         unsigned int s_mb_max_inode_prealloc;
  14 @@ -1536,6 +1539,9 @@
  15         atomic_t s_bal_goals;   /* goal hits */
  16         atomic_t s_bal_breaks;  /* too long searches */
  17         atomic_t s_bal_2orders; /* 2^order hits */
  18 +       /* cX loop didn't find blocks */
  19 +       atomic64_t s_bal_cX_failed[4];
  20 +       atomic64_t s_bal_cX_skipped[3];
  21         spinlock_t s_bal_lock;
  22         unsigned long s_mb_buddies_generated;
  23         unsigned long long s_mb_generation_time;
  24 @@ -2820,6 +2826,7 @@
  25  /* mballoc.c */
  26  extern const struct proc_ops ext4_seq_prealloc_table_fops;
  27  extern const struct seq_operations ext4_mb_seq_groups_ops;
  28 +extern const struct proc_ops ext4_mb_seq_alloc_fops;
  29  extern const struct proc_ops ext4_seq_mb_last_group_fops;
  30  extern int ext4_mb_seq_last_start_seq_show(struct seq_file *m, void *v);
  31  extern long ext4_mb_stats;
  32 diff -ur a/fs/ext4/mballoc.c b/fs/ext4/mballoc.c
  33 --- a/fs/ext4/mballoc.c 2021-06-28 08:57:13.669382587 -0600
  34 +++ b/fs/ext4/mballoc.c 2021-06-28 08:58:52.436392458 -0600
  35 @@ -2237,6 +2237,20 @@
  36         return ret;
  37  }
  38
  39 +static u64 available_blocks_count(struct ext4_sb_info *sbi)
  40 +{
  41 +       ext4_fsblk_t resv_blocks;
  42 +       u64 bfree;
  43 +       struct ext4_super_block *es = sbi->s_es;
  44 +
  45 +       resv_blocks = EXT4_C2B(sbi, atomic64_read(&sbi->s_resv_clusters));
  46 +       bfree = percpu_counter_sum_positive(&sbi->s_freeclusters_counter) -
  47 +                percpu_counter_sum_positive(&sbi->s_dirtyclusters_counter);
  48 +
  49 +       bfree = EXT4_C2B(sbi, max_t(s64, bfree, 0));
  50 +       return bfree - (ext4_r_blocks_count(es) + resv_blocks);
  51 +}
  52 +
  53  static noinline_for_stack int
  54  ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
  55  {
  56 @@ -2246,6 +2260,7 @@
  57         struct ext4_sb_info *sbi;
  58         struct super_block *sb;
  59         struct ext4_buddy e4b;
  60 +       ext4_fsblk_t avail_blocks;
  61
  62         sb = ac->ac_sb;
  63         sbi = EXT4_SB(sb);
  64 @@ -2298,6 +2313,21 @@
  65
  66         /* Let's just scan groups to find more-less suitable blocks */
  67         cr = ac->ac_2order ? 0 : 1;
  68 +
  69 +       /* Choose what loop to pass based on disk fullness */
  70 +       avail_blocks = available_blocks_count(sbi) ;
  71 +
  72 +       if (avail_blocks < sbi->s_mb_c3_blocks) {
  73 +               cr = 3;
  74 +               atomic64_inc(&sbi->s_bal_cX_skipped[2]);
  75 +       } else if(avail_blocks < sbi->s_mb_c2_blocks) {
  76 +               cr = 2;
  77 +               atomic64_inc(&sbi->s_bal_cX_skipped[1]);
  78 +       } else if(avail_blocks < sbi->s_mb_c1_blocks) {
  79 +               cr = 1;
  80 +               atomic64_inc(&sbi->s_bal_cX_skipped[0]);
  81 +       }
  82 +
  83         /*
  84          * cr == 0 try to get exact allocation,
  85          * cr == 3  try to get anything
  86 @@ -2361,6 +2391,9 @@
  87                         if (ac->ac_status != AC_STATUS_CONTINUE)
  88                                 break;
  89                 }
  90 +               /* Processed all groups and haven't found blocks */
  91 +               if (i == ngroups)
  92 +                       atomic64_inc(&sbi->s_bal_cX_failed[cr]);
  93         }
  94
  95         if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
  96 @@ -2643,6 +2676,95 @@
  97         .proc_write     = ext4_mb_last_group_write,
  98  };
  99
 100 +static int mb_seq_alloc_show(struct seq_file *seq, void *v)
 101 +{
 102 +       struct super_block *sb = seq->private;
 103 +       struct ext4_sb_info *sbi = EXT4_SB(sb);
 104 +
 105 +       seq_printf(seq, "mballoc:\n");
 106 +       seq_printf(seq, "\tblocks: %u\n", atomic_read(&sbi->s_bal_allocated));
 107 +       seq_printf(seq, "\treqs: %u\n", atomic_read(&sbi->s_bal_reqs));
 108 +       seq_printf(seq, "\tsuccess: %u\n", atomic_read(&sbi->s_bal_success));
 109 +
 110 +       seq_printf(seq, "\textents_scanned: %u\n",
 111 +                  atomic_read(&sbi->s_bal_ex_scanned));
 112 +       seq_printf(seq, "\t\tgoal_hits: %u\n", atomic_read(&sbi->s_bal_goals));
 113 +       seq_printf(seq, "\t\t2^n_hits: %u\n", atomic_read(&sbi->s_bal_2orders));
 114 +       seq_printf(seq, "\t\tbreaks: %u\n", atomic_read(&sbi->s_bal_breaks));
 115 +       seq_printf(seq, "\t\tlost: %u\n", atomic_read(&sbi->s_mb_lost_chunks));
 116 +
 117 +       seq_printf(seq, "\tuseless_c0_loops: %llu\n",
 118 +                  atomic64_read(&sbi->s_bal_cX_failed[0]));
 119 +       seq_printf(seq, "\tuseless_c1_loops: %llu\n",
 120 +                  atomic64_read(&sbi->s_bal_cX_failed[1]));
 121 +       seq_printf(seq, "\tuseless_c2_loops: %llu\n",
 122 +                  atomic64_read(&sbi->s_bal_cX_failed[2]));
 123 +       seq_printf(seq, "\tuseless_c3_loops: %llu\n",
 124 +                  atomic64_read(&sbi->s_bal_cX_failed[3]));
 125 +       seq_printf(seq, "\tskipped_c0_loops: %llu\n",
 126 +                  atomic64_read(&sbi->s_bal_cX_skipped[0]));
 127 +       seq_printf(seq, "\tskipped_c1_loops: %llu\n",
 128 +                  atomic64_read(&sbi->s_bal_cX_skipped[1]));
 129 +       seq_printf(seq, "\tskipped_c2_loops: %llu\n",
 130 +                  atomic64_read(&sbi->s_bal_cX_skipped[2]));
 131 +       seq_printf(seq, "\tbuddies_generated: %lu\n",
 132 +                  sbi->s_mb_buddies_generated);
 133 +       seq_printf(seq, "\tbuddies_time_used: %llu\n", sbi->s_mb_generation_time);
 134 +       seq_printf(seq, "\tpreallocated: %u\n",
 135 +                  atomic_read(&sbi->s_mb_preallocated));
 136 +       seq_printf(seq, "\tdiscarded: %u\n",
 137 +                  atomic_read(&sbi->s_mb_discarded));
 138 +       return 0;
 139 +}
 140 +
 141 +static ssize_t mb_seq_alloc_write(struct file *file,
 142 +                             const char __user *buf,
 143 +                             size_t cnt, loff_t *pos)
 144 +{
 145 +       struct ext4_sb_info *sbi = EXT4_SB(PDE_DATA(file_inode(file)));
 146 +
 147 +       atomic_set(&sbi->s_bal_allocated, 0),
 148 +       atomic_set(&sbi->s_bal_reqs, 0),
 149 +       atomic_set(&sbi->s_bal_success, 0);
 150 +
 151 +       atomic_set(&sbi->s_bal_ex_scanned, 0),
 152 +       atomic_set(&sbi->s_bal_goals, 0),
 153 +       atomic_set(&sbi->s_bal_2orders, 0),
 154 +       atomic_set(&sbi->s_bal_breaks, 0),
 155 +       atomic_set(&sbi->s_mb_lost_chunks, 0);
 156 +
 157 +       atomic64_set(&sbi->s_bal_cX_failed[0], 0),
 158 +       atomic64_set(&sbi->s_bal_cX_failed[1], 0),
 159 +       atomic64_set(&sbi->s_bal_cX_failed[2], 0);
 160 +       atomic64_set(&sbi->s_bal_cX_failed[3], 0);
 161 +
 162 +       atomic64_set(&sbi->s_bal_cX_skipped[0], 0),
 163 +       atomic64_set(&sbi->s_bal_cX_skipped[1], 0),
 164 +       atomic64_set(&sbi->s_bal_cX_skipped[2], 0);
 165 +
 166 +
 167 +       sbi->s_mb_buddies_generated = 0;
 168 +       sbi->s_mb_generation_time = 0;
 169 +
 170 +       atomic_set(&sbi->s_mb_preallocated, 0),
 171 +       atomic_set(&sbi->s_mb_discarded, 0);
 172 +
 173 +       return cnt;
 174 +}
 175 +
 176 +static int mb_seq_alloc_open(struct inode *inode, struct file *file)
 177 +{
 178 +       return single_open(file, mb_seq_alloc_show, PDE_DATA(inode));
 179 +}
 180 +
 181 +const struct proc_ops ext4_mb_seq_alloc_fops = {
 182 +       .proc_open      = mb_seq_alloc_open,
 183 +       .proc_read      = seq_read,
 184 +       .proc_lseek     = seq_lseek,
 185 +       .proc_release   = single_release,
 186 +       .proc_write     = mb_seq_alloc_write,
 187 +};
 188 +
 189  int ext4_mb_seq_last_start_seq_show(struct seq_file *m, void *v)
 190  {
 191         struct ext4_sb_info *sbi = EXT4_SB(m->private);
 192 @@ -2869,6 +2988,7 @@
 193         return 0;
 194  }
 195
 196 +#define THRESHOLD_BLOCKS(ts) (ext4_blocks_count(sbi->s_es) / 100 * ts)
 197  int ext4_mb_init(struct super_block *sb)
 198  {
 199         struct ext4_sb_info *sbi = EXT4_SB(sb);
 200 @@ -2923,6 +3043,9 @@
 201         sbi->s_mb_stats = MB_DEFAULT_STATS;
 202         sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
 203         sbi->s_mb_max_inode_prealloc = MB_DEFAULT_MAX_INODE_PREALLOC;
 204 +       sbi->s_mb_c1_blocks = THRESHOLD_BLOCKS(MB_DEFAULT_C1_THRESHOLD);
 205 +       sbi->s_mb_c2_blocks = THRESHOLD_BLOCKS(MB_DEFAULT_C2_THRESHOLD);
 206 +       sbi->s_mb_c3_blocks = THRESHOLD_BLOCKS(MB_DEFAULT_C3_THRESHOLD);
 207         /*
 208          * The default group preallocation is 512, which for 4k block
 209          * sizes translates to 2 megabytes.  However for bigalloc file
 210 @@ -3062,6 +3185,17 @@
 211                                 atomic_read(&sbi->s_bal_reqs),
 212                                 atomic_read(&sbi->s_bal_success));
 213                 ext4_msg(sb, KERN_INFO,
 214 +                       "mballoc: (%llu, %llu, %llu, %llu) useless c(0,1,2,3) loops",
 215 +                               atomic64_read(&sbi->s_bal_cX_failed[0]),
 216 +                               atomic64_read(&sbi->s_bal_cX_failed[1]),
 217 +                               atomic64_read(&sbi->s_bal_cX_failed[2]),
 218 +                               atomic64_read(&sbi->s_bal_cX_failed[3]));
 219 +               ext4_msg(sb, KERN_INFO,
 220 +                       "mballoc: (%llu, %llu, %llu) skipped c(0,1,2) loops",
 221 +                               atomic64_read(&sbi->s_bal_cX_skipped[0]),
 222 +                               atomic64_read(&sbi->s_bal_cX_skipped[1]),
 223 +                               atomic64_read(&sbi->s_bal_cX_skipped[2]));
 224 +               ext4_msg(sb, KERN_INFO,
 225                       "mballoc: %u extents scanned, %u goal hits, "
 226                                 "%u 2^N hits, %u breaks, %u lost",
 227                                 atomic_read(&sbi->s_bal_ex_scanned),
 228 diff -ur a/fs/ext4/mballoc.h b/fs/ext4/mballoc.h
 229 --- a/fs/ext4/mballoc.h 2021-06-28 08:57:13.337385973 -0600
 230 +++ b/fs/ext4/mballoc.h 2021-06-28 08:58:52.436392458 -0600
 231 @@ -68,6 +68,9 @@
 232   * for which requests use 2^N search using buddies
 233   */
 234  #define MB_DEFAULT_ORDER2_REQS         8
 235 +#define MB_DEFAULT_C1_THRESHOLD                25
 236 +#define MB_DEFAULT_C2_THRESHOLD                15
 237 +#define MB_DEFAULT_C3_THRESHOLD                5
 238
 239  /*
 240   * default group prealloc size 512 blocks
 241 diff -ur a/fs/ext4/sysfs.c b/fs/ext4/sysfs.c
 242 --- a/fs/ext4/sysfs.c   2021-06-28 08:57:13.669382587 -0600
 243 +++ b/fs/ext4/sysfs.c   2021-06-28 09:00:20.199538001 -0600
 244 @@ -21,6 +21,9 @@
 245  typedef enum {
 246         attr_noop,
 247         attr_delayed_allocation_blocks,
 248 +       attr_mb_c1_threshold,
 249 +       attr_mb_c2_threshold,
 250 +       attr_mb_c3_threshold,
 251         attr_session_write_kbytes,
 252         attr_lifetime_write_kbytes,
 253         attr_reserved_clusters,
 254 @@ -140,6 +143,32 @@
 255                         task_pid_vnr(sbi->s_journal->j_task));
 256  }
 257
 258 +#define THRESHOLD_PERCENT(ts) (ts * 100 / ext4_blocks_count(sbi->s_es))
 259 +
 260 +static int save_threshold_percent(struct ext4_sb_info *sbi, const char *buf,
 261 +                                 ext4_fsblk_t *blocks)
 262 +{
 263 +       unsigned long long val;
 264 +
 265 +       int ret;
 266 +
 267 +       ret = kstrtoull(skip_spaces(buf), 0, &val);
 268 +       if (ret || val > 100)
 269 +               return -EINVAL;
 270 +
 271 +       *blocks = val * ext4_blocks_count(sbi->s_es) / 100;
 272 +       return 0;
 273 +}
 274 +
 275 +static ssize_t mb_threshold_store(struct ext4_sb_info *sbi,
 276 +                                 const char *buf, size_t count,
 277 +                                 ext4_fsblk_t *blocks)
 278 +{
 279 +       int ret = save_threshold_percent(sbi, buf, blocks);
 280 +
 281 +       return ret ?: count;
 282 +}
 283 +
 284  #define EXT4_ATTR(_name,_mode,_id)                                     \
 285  static struct ext4_attr ext4_attr_##_name = {                          \
 286         .attr = {.name = __stringify(_name), .mode = _mode },           \
 287 @@ -202,6 +231,9 @@
 288  #define ATTR_LIST(name) &ext4_attr_##name.attr
 289
 290  EXT4_ATTR_FUNC(delayed_allocation_blocks, 0444);
 291 +EXT4_ATTR_FUNC(mb_c1_threshold, 0644);
 292 +EXT4_ATTR_FUNC(mb_c2_threshold, 0644);
 293 +EXT4_ATTR_FUNC(mb_c3_threshold, 0644);
 294  EXT4_ATTR_FUNC(session_write_kbytes, 0444);
 295  EXT4_ATTR_FUNC(lifetime_write_kbytes, 0444);
 296  EXT4_ATTR_FUNC(reserved_clusters, 0644);
 297 @@ -251,6 +283,9 @@
 298
 299  static struct attribute *ext4_attrs[] = {
 300         ATTR_LIST(delayed_allocation_blocks),
 301 +       ATTR_LIST(mb_c1_threshold),
 302 +       ATTR_LIST(mb_c2_threshold),
 303 +       ATTR_LIST(mb_c3_threshold),
 304         ATTR_LIST(session_write_kbytes),
 305         ATTR_LIST(lifetime_write_kbytes),
 306         ATTR_LIST(reserved_clusters),
 307 @@ -367,6 +402,15 @@
 308                 return snprintf(buf, PAGE_SIZE, "%llu\n",
 309                                 (s64) EXT4_C2B(sbi,
 310                        percpu_counter_sum(&sbi->s_dirtyclusters_counter)));
 311 +       case attr_mb_c1_threshold:
 312 +               return scnprintf(buf, PAGE_SIZE, "%llu\n",
 313 +                                THRESHOLD_PERCENT(sbi->s_mb_c1_blocks));
 314 +       case attr_mb_c2_threshold:
 315 +               return scnprintf(buf, PAGE_SIZE, "%llu\n",
 316 +                                THRESHOLD_PERCENT(sbi->s_mb_c2_blocks));
 317 +       case attr_mb_c3_threshold:
 318 +               return scnprintf(buf, PAGE_SIZE, "%llu\n",
 319 +                                THRESHOLD_PERCENT(sbi->s_mb_c3_blocks));
 320         case attr_session_write_kbytes:
 321                 return session_write_kbytes_show(sbi, buf);
 322         case attr_lifetime_write_kbytes:
 323 @@ -468,6 +512,12 @@
 324                 return inode_readahead_blks_store(sbi, buf, len);
 325         case attr_trigger_test_error:
 326                 return trigger_test_error(sbi, buf, len);
 327 +       case attr_mb_c1_threshold:
 328 +               return mb_threshold_store(sbi, buf, len, &sbi->s_mb_c1_blocks);
 329 +       case attr_mb_c2_threshold:
 330 +               return mb_threshold_store(sbi, buf, len, &sbi->s_mb_c2_blocks);
 331 +       case attr_mb_c3_threshold:
 332 +               return mb_threshold_store(sbi, buf, len, &sbi->s_mb_c3_blocks);
 333         }
 334         return 0;
 335  }
 336 @@ -530,6 +580,8 @@
 337                                 &ext4_seq_mb_last_group_fops, sb);
 338                 proc_create_single_data("mb_last_start", S_IRUGO, sbi->s_proc,
 339                                 ext4_mb_seq_last_start_seq_show, sb);
 340 +               proc_create_data("mb_alloc_stats", S_IFREG | S_IRUGO | S_IWUSR,
 341 +                                sbi->s_proc, &ext4_mb_seq_alloc_fops, sb);
 342         }
 343         return 0;
 344  }